Claude Fable 5.1 tops the Artificial Analysis Intelligence Index -- and costs 20% more per task than Fable 5 despite a 75% cache-price cut

Anthropic's newest flagship, Claude Fable 5.1, now holds the top spot on Artificial Analysis's Intelligence Index -- the independent, cross-lab benchmark behind every score on the Scoreboard. At its max-effort setting the model scores 66, four points ahead of Anthropic's own Claude Opus 5 (63) and two ahead of GPT-5.6 Sol and Grok 4.6 (61 each). The release, which shipped September 1 alongside a restricted-access twin called Claude Mythos 5.1, also cut the price of Anthropic's most heavily used pricing lever -- cached context -- by three-quarters. Those two facts point in different directions on the question every buyer actually asks: is this model worth more, or does it simply cost less?

The 66 figure is Fable 5.1's ceiling, not its default. Anthropic ships the model across five effort tiers, and Artificial Analysis measured all of them: 58 at the lowest setting, climbing to 65 at "xhigh" and 66 at max -- a span that costs 13.1 million output tokens at the low end and 143.7 million at the top, across the benchmark's full test suite. That's an 11-fold difference in tokens spent for an 8-point swing in score, which is a real tradeoff a budget-conscious buyer has to actually choose between, not a single number to quote.

That ranking isn't only Artificial Analysis's read. Cursor, the coding-agent tool that added Fable 5.1 to its model picker on release day, ran its own multi-file coding test -- CursorBench 3.2 -- and landed on the same order: Fable 5.1 at 73.4%, ahead of Fable 5's 70.5% and Claude Opus 5's 70.0%. Cursor's team pointed to a specific behavioral change behind the number rather than raw scale: Fable 5.1 catches and corrects its own mistakes mid-run on long, unattended coding sessions, instead of compounding an early error across the rest of the task.

The pricing change is where Anthropic's own announcement puts the emphasis. Base rates hold steady at $10 per million input tokens and $50 per million output tokens -- unchanged from Fable 5 -- but the price of a cache read, the charge applied when the model reuses context it has already processed, dropped from $1.00 to $0.25 per million tokens. Anthropic says that cuts the cost of a typical workload by about 25%, and a heavily agentic one -- the kind that revisits the same codebase, system prompt, and tool definitions turn after turn -- by up to 45%.

Cache reads with Fable 5.1 cost 75% less than Fable 5's. This reduces the cost of the model in practice by around 25% for typical workloads, and up to 45% for highly agentic ones.

Read only that far and Fable 5.1 looks like a straightforward win: a higher score at a lower price. Artificial Analysis's own task-level cost measurement says otherwise. At max effort, running Fable 5.1 through the benchmark's test suite cost $3.76 per task -- about 20% more than Fable 5's $3.14, and well above Claude Opus 5's $2.34. The gap isn't a pricing error; it's a tokens problem. Fable 5.1 at max effort uses roughly 1.7 times the output tokens Fable 5 needed to reach its lower score, and the cache-read discount -- worth an estimated $1.40 off that same task, per Artificial Analysis -- isn't large enough to offset the extra generation.

The other half of the release is priced identically but restricted: Claude Mythos 5.1. Same underlying weights, fewer restrictions -- available only to vetted organizations in Anthropic's Cyber Verification Program and Life Sciences Verification Program, both currently limited to US organizations. Anthropic says the cybersecurity variant now throws 60% fewer false positives when flagging misuse than its predecessor, and can identify a vulnerability without generating a working exploit for it. It amounts to Anthropic publishing two safety postures for the same underlying capability and letting the access program, not one global setting, decide which a given user gets.

Fable 5.1 is Anthropic's fourth model refresh since Opus 5 became the flagship default on July 24, following Mythos 5, Fable 5, and Sonnet 5 in June -- a cadence of roughly one release every five to six weeks through the summer. (The effort-tier system itself -- five settings trading tokens for score -- isn't unique to Anthropic; OpenAI's GPT-5.6 line ships a comparable low/high/max split, and Grok 4.6 offers its own "high" tier.) What's new this time isn't the capability jump on its own -- four points on the index is in line with recent generational steps -- it's Anthropic explicitly pricing the gap between "cheaper" and "better" instead of folding both into one release note.

None of this makes Fable 5.1 a bad release -- it's the highest score Artificial Analysis has measured on any model, full stop, and the cache-pricing change is a genuine cut for the workloads it targets. But the model that tops a leaderboard and the model that costs less to run are not, on Anthropic's own numbers, the same model this time -- and a buyer choosing between Fable 5.1 and its predecessor needs both figures, not just the one on the announcement page.